Efficient Assessment of Individualized Disease Risk and Treatment Response via Augmentation

Abstract

T-year survival, defined as the survival status by a pre-specified time point t, is of great interest in many medical research areas. When the t-year survival is the outcome of interest in the individualized medicine, baseline covariates are used to predict the t-year survival for potential treatment response comparison. Time-specific generalized linear models estimated with inverse censoring probability weighting provides more robustness to model misspecification compared to other methods, but some challenges remain in the heavy censoring settings: the prediction model could be quite inefficient and deriving the optimal individualized treatment rules based on maximizing the population average survival probability could be difficult. Chapter 1 presents an imputation-based method to improve the efficiency of the baseline prediction model by incorporating the information from subjects censored before t and auxiliary covariates including the post-baseline secondary outcomes collected before censoring. Chapter 2 extends the method in Chapter 1 to incorporate the post-baseline intermediate covariates that are collected before t but have non-negligible missing values due to censoring. Chapter 3 proposes a systematic approach to derive optimal individualized treatment rules that maximizes the population average survival probability, and imputation-based augmentation approach is also developed to improve the efficiency of the estimation.